Abstract: A common problem encountered in high frequency financial time series is the occurrence of extreme observations, or significant spikes in volatility, with subsequent influence on model specification, parameter estimation and future predictions in ARMA and GARCH models. We present estimated biases of Garch(1,1) model coefficients by unrecognized exogenous interventions at unknown dates with particular attention to additive level outliers. We further determine their approximate and simulated influences on estimated predictions through the inflation of innovation variance estimates. Our conclusion maintains that the severity of bias depends on the distance of an outlier from the prediction origin.